diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/layer.py b/python/sglang/srt/layers/moe/fused_moe_triton/layer.py
index aad266166..544c0494e 100644
--- a/python/sglang/srt/layers/moe/fused_moe_triton/layer.py
+++ b/python/sglang/srt/layers/moe/fused_moe_triton/layer.py
@@ -68,7 +68,6 @@ from sglang.srt.utils import (
     get_bool_env_var,
     is_cpu,
     is_hip,
-    is_npu,
     print_info_once,
     round_up,
 )
@@ -77,7 +76,6 @@ from sglang.srt.utils.custom_op import register_custom_op
 _is_hip = is_hip()
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_cpu = is_cpu()
-_is_npu = is_npu()
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
 
@@ -791,13 +789,6 @@ class FusedMoE(torch.nn.Module):
         # expert weights into block layout. During weight update, we must restore
         # canonical load-time shapes before copying checkpoint tensors.
         if isinstance(method, UnquantizedFusedMoEMethod):
-            if _is_npu:
-                if weight_name.endswith(".experts.w2_weight"):
-                    if param.data.shape[1] != loaded_weight.shape[0]:
-                        param.data = param.data.transpose(1, 2).contiguous()
-                if weight_name.endswith(".experts.w13_weight"):
-                    if param.data.shape[2] != loaded_weight.shape[1]:
-                        param.data = param.data.transpose(1, 2).contiguous()
             method.maybe_restore_flashinfer_trtllm_bf16_weight_shape_for_load(
                 layer=self,
                 param=param,
diff --git a/python/sglang/srt/layers/quantization/unquant.py b/python/sglang/srt/layers/quantization/unquant.py
index efd3d807b..fe2d140ff 100644
--- a/python/sglang/srt/layers/quantization/unquant.py
+++ b/python/sglang/srt/layers/quantization/unquant.py
@@ -337,14 +337,10 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
             layer.w2_weight.data = layer.w2_weight.data.reshape(
                 layer.num_local_experts, *new_shape_w2
             )
-
         if _is_npu:
             for weight_name in ["w13_weight", "w2_weight"]:
                 weight = getattr(layer, weight_name)
-                origin_weight = weight.data.transpose(1, 2)
-                new_weight = origin_weight.contiguous()
-                origin_weight.untyped_storage().resize_(0)
-                weight.data = npu_format_cast(new_weight)
+                weight.data = npu_format_cast(weight)
 
         return
 
@@ -697,7 +693,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
         # gmm1: gate_up_proj
         hidden_states = torch.ops.npu.npu_grouped_matmul(
             x=[hidden_states],
-            weight=[layer.w13_weight],
+            weight=[layer.w13_weight.transpose(1, 2)],
             bias=w13_bias,
             split_item=2,
             group_list_type=1,
@@ -733,7 +729,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
         # gmm2: down_proj
         hidden_states = torch.ops.npu.npu_grouped_matmul(
             x=[hidden_states],
-            weight=[layer.w2_weight],
+            weight=[layer.w2_weight.transpose(1, 2)],
             bias=w2_bias,
             split_item=2,
             group_list_type=1,